Unverified50% confidenceBenchmarkExact time
Under the four-reward setting, WAS-PO achieved a higher average training reward across all three random seeds, by 3.9%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/3/2026
First Seen
Valid until: 2/1/2027
Sources
Related Entities
Related Claims
UnverifiedTraining hyperparameters: one PPO epoch per update, Adam learning rate 3e-4, discount factor γ=0.9997 (about 7 minutes of horizon, consistent with OpenAI Five), and a win reward of ±10.66% similarUnverifiedOnce GOOSE is included in the training set, all augmentation branches perform within ±1 point of the control group on real weather, while augmentation still adds 3.5 to 5.2 points on synthetic held-out degradations.61% similarUnverifiedThe 3.11x is kernel-level forward acceleration, while 34% is end-to-end training step speed improvement; the two are measured on different bases.60% similarUnverifiedGPT-3 training used the Adam optimizer (beta1 = 0.9, beta2 = 0.95), gradient clipping to 1.0, a learning rate warmed up over 375 million tokens then cosine-decayed to 10% of its peak over 260 billion tokens, and weight decay of 0.1.60% similarUnverifiedSkill-Uni was tested on five skill-related benchmarks, improving on average by 66.59% over Qwen3-4B and by 34.82% on average over LLaMA3.1-8B.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130537API
curl https://kongchang.com/api/v1/knowledge/claims/1130537MCP
get_claim(id=1130537)